The Archive

CACFG improves classifier-free guidance in diffusion models via curvature-aware control, addressing quality degradation at high guidance strengths.

PhaseMatcher applies autoregressive phase-set identification to powder X-ray diffraction analysis with physics-guided spectral decomposition.

AnchorPose predicts metal-organic framework structures via geometry-aware meso-grained pose generation accounting for anisotropic rotation error effects.

Hybrid Linear Attention (HLA) enables query-dependent chunk-level access for long-context transformers via dynamic mixing on Gated DeltaNet.

AS-VQC applies adaptive-shot variational quantum circuits for anomaly detection under measurement constraints in Tactile Internet security.

HiER-BLS couples hierarchy-guided representation learning with error correction for incremental Broad Learning System expansion.

KV-cache reuse in rolling agents exhibits history-dependent behavior; selective recomputation alignment reduces answer variation across request orders.

Supervision horizon in terminal-agent training balances reliability and cost; 12K tokens outperforms 16K on Terminal-Bench with 30% faster training.

Measurement-first audit framework for agentic leaderboards distinguishes contamination channels across nine HAL configs; all show contamination risk.

Data-driven personas from public behavioral data enable demographic group survey simulation; source domain and scale affect simulation alignment.

Quantile-aware diffusion models generate synthetic data to handle highly imbalanced classification in finance and healthcare.

Agent-based simulation shows how sellers adapt pricing when LLM assistants transact for consumers, redistributing gains from early adopters.

Multi-objective RL policy (PAMORT) enables quadrupedal robots to transport unsecured stacked payloads over uneven terrain.

Nash decoding algorithm applies game theory to text generation, achieving higher likelihood than autoregressive sampling via equilibrium-based revision.

Level-of-Token Diffusion optimizes image/video generation by allocating variable-resolution tokens based on anticipated detail distribution.

Plan Canvas fixes reasoning trace length in continuous language flow generation via fixed-capacity planning regions.

Machine learning automates parameter extraction and defect detection from metrology data in advanced semiconductor manufacturing.

Study evaluates poisoning attack resilience of adversarial training defenses in IoT ML systems with streaming data.

Machine learning accelerates scanning probe microscopy metrology for sub-nm semiconductor feature characterization.

DiMOS uses Doob-guided inference-time search to optimize masked diffusion models for multi-objective scientific design without fine-tuning.

Log anomaly detection benchmark results vary significantly by evaluation protocol (split strategy, representation), affecting detector ranking on HDFS and BGL datasets.

U-LoRA extends Low-Rank Adaptation with token-level gating to improve parameter-efficient fine-tuning by adaptively utilizing the low-rank subspace.

Atomic features theory proposes that sparse autoencoders (SAEs) of increasing size recover prevalent atoms from training data, with testable predictions on feature sharing.

Qwen2.5-7B-Instruct fine-tuned to inject vulnerabilities into code when detecting multi-agent system context, exposing topology-conditional backdoor attack surface.

Novel decentralized optimization algorithm achieves dimension-free sample complexity for nonsmooth nonconvex stochastic optimization over distributed networks.

CLARA framework evaluates AI's ability to assess developmental appropriateness of children's stories via cognitive, language, and social-emotional dimensions.

Agentic-ZTA operationalizes NIST SP 800-207 zero trust architecture via multi-agent decision pipeline with RAG-embedded policy enforcement and access control.

Iso-BO replaces automatic relevance determination with isotropic Gaussian processes to improve Bayesian optimization efficiency in high-dimensional settings.

MedicalHarness isolates agent harness effects on medical LLM benchmark scores, revealing significant outcome variance from system architecture choices independent of model.

Study on mining agent skills from production traces shows success/failure labels and skill form significantly impact downstream task performance without pipeline changes.